Measuring impact when the click disappears — zero-click value and assisted conversions.
Attribution assumes a click. AI search removes it: your brand can be named, described, and recommended inside an answer that the user never leaves. Last-click models record nothing for these influenced decisions, which means the more your visibility shifts into answers, the more your reporting understates it. The workable response is not to invent a click that did not happen, but to measure AI citations the way brand and PR have always been measured — by influence, share, and the demand they create.
AI search is substantially zero-click: the engine composes an answer, cites its sources, and the user gets what they needed without visiting any of them. Your content can shape the answer — your brand recommended, your data quoted, your position represented — while your analytics records no session at all. The influence is real and the visit is absent.
This is not a measurement bug but a structural change in how information reaches people. Traffic-based measurement was always a proxy for influence, and it worked while influence reliably produced visits. As answers absorb more of that journey, the proxy degrades. Understanding the zero-click problem correctly is the first step: the value did not disappear, the click did, and measurement has to follow the value rather than the click.
Last-click attribution assigns credit to the final touchpoint before conversion, which works reasonably when the journey is a chain of clicks. AI citations sit outside that chain entirely. Someone who asks an assistant which vendors to consider, receives an answer citing you, and later searches your brand name directly will have that conversion attributed to branded search — or to direct traffic — with the citation that created the demand recorded nowhere.
The result is a systematic bias: channels that produce clicks get credit, while channels that produce influence without clicks appear worthless. Under a last-click model, an AI-visibility programme that is materially growing demand can look like it is contributing nothing. Understanding this failure mode matters because it leads directly to defunding the work that is actually working, which is the practical danger of unexamined attribution.
The most useful mental shift is to stop treating AI citations as a traffic channel and start treating them as brand exposure. Public relations has never been measured by last-click, because nobody expects a favourable mention in a publication to produce an immediately traceable conversion. It is measured by reach, share of coverage, sentiment, and the demand that follows.
AI citations fit that model well: they are third-party mentions in a trusted context, reaching people at a decision-relevant moment, whose effect shows up as demand rather than as a session. Adopting the brand-and-PR frame gives you an established measurement vocabulary rather than requiring a new one. Understanding this reframing resolves most of the confusion, because the measurement problem is not novel — it is the one brand marketing has always had.
The most direct measure of AI-search performance is citation share: for the questions that matter to your business, how often are you cited, and how does that compare with competitors. It is the closest analogue to share of voice, and it functions as a leading indicator — it moves before the downstream demand effects appear, which makes it useful for steering.
Its virtue is that it measures the thing you actually control: whether your content is being selected as a source. Unlike downstream demand signals, it is not confounded by every other marketing activity. Understanding citation share as the leading indicator is why AI-visibility measurement should start there, with demand signals used to corroborate that the citations are producing effects rather than to establish whether citations are occurring.
The most accessible downstream signal is branded search volume. If AI answers are exposing your brand to people researching your category, some proportion will subsequently search for you by name. Branded search growth — particularly when it rises without a corresponding increase in other brand-building activity — is reasonable evidence that upstream exposure is creating demand.
Related signals include direct traffic growth and increases in unattributed conversions from people arriving already aware of you. None is conclusive alone, since many things drive brand awareness, but together and tracked against citation share over time they form a coherent picture. Understanding branded search as a corroborating signal is why it belongs alongside citation tracking: one shows the exposure, the other shows the demand that follows it.
Last-click records nothing for an answer the user never left. Measure the exposure you control — citation share against competitors — and corroborate it with branded search, direct traffic, and informed conversions.
Between exposure and revenue sits a category worth measuring: conversions from people who arrived already informed. These show up as shorter consideration cycles, fewer pre-purchase questions, higher conversion rates from branded and direct traffic, and prospects who reference specifics they could only have learned before contacting you.
Sales conversations and post-purchase surveys are unusually informative here, because asking customers how they first heard of you frequently surfaces AI assistants explicitly — evidence no analytics platform will provide. The practical method is to add that question and track the answers. Understanding informed conversions as a measurable category is why qualitative evidence deserves a place in this reporting: it captures influence that instrumentation structurally cannot.
The analytical move that makes this credible is correlation over time. Track citation share and downstream demand signals on the same timeline, and look for whether movements in one precede movements in the other. If citation share grows in a set of topics and branded search and informed conversions rise afterwards, the relationship is at least plausible.
This is correlational rather than causal, and it should be presented as such. But it is considerably better than either claiming precise attribution you cannot support or reporting nothing at all. Understanding how to connect citations to demand over time is what turns AI-visibility measurement from an assertion of value into an evidenced argument, which is what stakeholders reasonably require.
You can’t attribute an answer nobody clicked — but you can measure whether you’re in it. DUNkē tracks citation share across eight AI engines, per prompt and against competitors, so the influence last-click misses becomes a number you can trend.
Credibility depends on not overclaiming. Do not assert that a specific citation produced a specific conversion, because you cannot know that. Do not construct attribution models that assign precise revenue to AI citations on assumptions the data does not support. And do not present correlation as proof of causation, because a sceptical stakeholder will find the weakness and discount everything else you presented alongside it.
The stronger position is honest: we are cited this often for these questions, our share against competitors is this, and here is the demand trend that has followed. That is a defensible argument. Understanding what not to claim is what protects the credibility of the whole measurement programme, since a single indefensible claim tends to undermine the reasonable ones surrounding it.
Presenting influence-based measurement to people accustomed to click-based reporting requires framing. The move that works is to establish the structural point first — that AI answers deliver visibility without visits, and that click-based reporting therefore understates them by design — before presenting the alternative measures. Without that framing, citation share reads as a substitute metric offered in place of results.
The brand-and-PR analogy carries most of the weight, because it is familiar and the reasoning is accepted: nobody demands last-click attribution for coverage in a major publication. Understanding how to frame this is why the reporting discipline matters as much as the measurement — the numbers only persuade if the reason for using them has been established first.
The recurring mistakes are consequential. Judging AI-search work by last-click performance guarantees it appears worthless. Ignoring the zero-click shift entirely leaves reporting drifting further from reality as answers grow. Overclaiming precise attribution destroys credibility. Measuring only downstream demand without tracking citation share leaves you unable to tell whether your own actions caused the movement. And never asking customers how they found you discards the most direct evidence available.
The remedies are to adopt the brand-and-PR frame explicitly, track citation share as the controllable leading indicator, corroborate with branded search, direct traffic, and informed conversions, present correlation honestly, and ask customers directly. Understanding these failure modes matters because attribution errors here do not merely misreport — they actively misdirect investment away from work that is producing results.
It helps to recognise that traffic-based measurement was never measuring value directly — it was measuring a convenient correlate. The value of search visibility was always the influence on someone’s decision; the click was simply the observable event that reliably accompanied it, and instrumentation made it cheap to count.
Seeing the click as a proxy rather than the thing itself reframes the current problem accurately. The proxy is degrading because influence increasingly occurs without a visit, not because value is disappearing. Understanding this is what makes the shift to influence-based measurement feel like a correction rather than a concession — you are moving closer to what you always wanted to measure, having lost the shortcut that used to stand in for it.
Even before AI answers, most decisions involved many touchpoints — searches, articles, recommendations, conversations, advertisements — of which analytics observed only a fraction. Attribution models were always allocating credit across an incomplete picture, and the confidence with which they reported was always somewhat overstated.
AI citations extend an existing problem rather than creating a new one. This matters for how you present the shift: you are not asking stakeholders to abandon reliable measurement for something vaguer, but pointing out that a known limitation has grown large enough to require a response. Understanding the multi-touch reality is a useful framing, because it positions influence measurement as maturity rather than as an excuse for unmeasurable work.
The most direct evidence about AI-search influence comes from asking. A simple question at the point of conversion or in onboarding — how did you first hear about us, in an open field rather than a fixed list — regularly surfaces mentions of AI assistants that no analytics platform records.
Self-reported data has known weaknesses: people misremember, and answers skew toward memorable touchpoints. But it captures influence that instrumentation structurally cannot, and at scale the pattern is informative even if individual responses are imperfect. The practical step is to add the question and track responses over time. Understanding self-reported attribution is why the lowest-technology method available is often the most revealing one here.
If AI citations create demand that surfaces as branded search, then branded traffic deserves closer examination than a single aggregate line. Splitting it by query type — pure brand name, brand plus product, brand plus a specific question — reveals different awareness states, and growth in the more specific variants often indicates people arriving with knowledge acquired elsewhere.
Tracking new versus returning visitors within branded traffic adds another dimension, since growth in new branded visitors suggests genuinely new awareness rather than existing customers returning. The practical method is to build these splits into standing reporting. Understanding branded segmentation is why the demand signal is more informative than it first appears, provided you look inside the aggregate.
Where circumstances allow, comparison strengthens the argument considerably. If you can improve citation share substantially for one set of topics while leaving another comparable set unchanged, and the demand signals for the first move while the second does not, you have evidence approaching a controlled comparison.
This is rarely clean in practice, since other activity interferes and topics differ. But even an imperfect comparison is more persuasive than a single trend line, because it addresses the obvious objection that something else caused the movement. Understanding the value of controlled comparison is why staged rollouts across topic areas are worth structuring deliberately when you have the opportunity.
Faced with genuine attribution difficulty, some organisations conclude that AI-search influence cannot be measured and therefore leave it unmeasured. This is the worst available option, because unmeasured work is unfunded work: budget flows to the channels that report numbers, regardless of where value is actually created.
Imperfect measurement, presented honestly with its limitations stated, is considerably better than none. Citation share plus corroborating demand signals is a defensible position; silence is not. Understanding this risk is why the practical priority is to establish some credible measurement quickly rather than to wait for a rigorous solution, since the absence of any number is itself a decision with consequences.
AI-citation influence often operates on a longer horizon than click-based channels. Someone who encounters your brand in an answer while researching may act weeks or months later, after further consideration — which means the effects appear outside the attribution windows most analytics configurations use.
The practical adjustment is to look at longer horizons when assessing whether citation growth has produced demand, and to be sceptical of short-window assessments that show nothing. A quarter is often the shortest sensible period. Understanding delayed effects is why judging AI-visibility work on monthly conversion data will usually and misleadingly show no effect, regardless of whether one exists.
The practical output of all this is a measurement story with several strands rather than one number: citation share and its trend, showing the exposure you have built; branded search and direct traffic trends, showing demand movement; informed-conversion indicators and self-reported attribution, showing customers arriving pre-educated; and the timeline connecting them.
Presented together with limitations stated plainly, this is genuinely persuasive, because each strand addresses a different objection and the convergence is harder to dismiss than any single measure. Understanding that the story has strands rather than a headline metric is why the reporting should be built as an argument, and why attempting to reduce it to one number invariably makes it weaker.
Attribution tooling for AI search is immature and will improve, which raises the question of how to operate now. The workable answer is to build the measurement habits that will remain valuable regardless: a stable prompt set with citation tracking, clean branded-traffic segmentation, a self-reported attribution question, and archived history so that trends exist when better analysis becomes possible.
These are inexpensive and compound, since the historical record you build now is what future analysis will need and cannot retrospectively create. Understanding what to do in the interim is why the right response to immature tooling is to start collecting rather than to wait — the data you fail to gather this year is the baseline you will wish you had next year.
The most damaging attribution arguments happen after the numbers are in, when a stakeholder disputes the measurement approach specifically because they dislike what it shows. The way to avoid this is to agree how AI-search work will be judged before it starts: which measures count, what constitutes progress, over what horizon, and what would indicate the programme is not working.
This is a straightforward conversation to have in advance and a fraught one to have retrospectively, since any measurement proposed after disappointing results looks like special pleading. Agreeing the model up front also forces useful clarity about what the investment is expected to achieve. Understanding the timing of this conversation is why the measurement framework belongs in the programme proposal rather than in the first review, and why deferring it tends to cost the programme its credibility later.
Zero-click is a tendency rather than an absolute: AI answers do link their cited sources, and some proportion of users click through, particularly for questions where they want detail the summary omitted or want to verify a claim. This traffic is worth identifying, because it is unusually well-qualified — these visitors arrived having already encountered you in a trusted context and chose to investigate further.
Identifying it requires attention, since referral data from assistants is inconsistent and often appears as direct traffic. Where it can be isolated, its conversion behaviour is worth examining separately rather than averaging it into general organic performance. Understanding that some traffic does arrive prevents the overcorrection of treating AI visibility as entirely trafficless, and the visitors it produces are frequently among the most valuable a site receives.
The attribution difficulty is ultimately organisational rather than technical. The measurement approaches described here are workable; what makes them hard to adopt is that budgets, incentives, and reporting structures are frequently built around click-based metrics, and changing what a team is measured by is a political undertaking rather than an analytical one.
This is worth naming, because teams often treat their attribution problem as a search for a better tool when the actual obstacle is that nobody senior has agreed that influence counts. The practical implication is that the most valuable work may be securing that agreement rather than refining the measurement. Understanding attribution as an organisational problem is why the argument for influence-based measurement has to be made to decision-makers, not merely documented in a dashboard nobody has authorised.
As influence-based measurement becomes established in an organisation, a subtle risk emerges: the measures start being managed rather than merely reported. Prompt sets drift toward questions where performance is strong, comparisons get made against periods that flatter, and the definition of a citation quietly loosens. None of this requires bad faith; it happens gradually whenever a number becomes something a team is judged by.
The protection is procedural rather than analytical: document the methodology explicitly, hold the prompt set stable except at scheduled reviews, record any change to how measures are defined, and note methodology changes alongside the trend so nobody reads a definitional shift as performance. Understanding this risk matters because influence-based measurement is more susceptible to it than click-based measurement, precisely because its definitions involve more judgment — which makes documented discipline the thing that keeps it credible.
AI search is zero-click by design: you can be cited, described, and recommended inside an answer the user never leaves, which means last-click attribution records nothing for decisions you demonstrably influenced. The consequence is systematic understatement — and the practical danger is defunding work that is actually creating demand because the model cannot see it.
The workable response is to treat AI citations as brand exposure rather than as a traffic channel, borrowing the measurement logic that public relations has always used. Lead with citation share, the controllable leading indicator; corroborate with branded search growth, direct traffic, and informed conversions; ask customers directly how they first encountered you; and connect citations to demand over time while presenting the relationship honestly as correlation. Measure the value, not the click that no longer happens.
“Nobody demands last-click attribution for a favourable mention in a major publication. An AI citation is the same kind of thing — measure the share you win and the demand that follows.” The Age’X Research Team
Get a free GEO audit — the same analysis behind every article here.